3, MARCH 2010
Receiver I/Q Imbalance: Tone Test, Sensitivity
Analysis, and the Universal Software
Radio Peripheral
Peter Hndel, Senior Member, IEEE, and Per Zetterberg
AbstractThe problem of determining the gain imbalance,
quadrature skew, and local oscillator leakage of contemporary ra
dio frequency receivers by tone test is considered. A least squares
approach for indirect estimation of the soughtfor parameters
is proposed, which is linear in six out of its seven parameters.
The performance of the method, particularly its accuracy as a
function of measurement time, imbalance parameters, and sig
naltonoise ratios, is investigated. The theoretical predictions of
the performance are illustrated by Monte Carlo simulations and
experimental data that are obtained from testing several universal
software radio peripheral (USRP) receivers. This paper shows
that, for the studied offtheshelf receivers, gain imbalance and
quadrature skew may accurately be predicted (i.e., < 0.1 dB and
< 1
o
n +
I
+
2
_
+c
I
(5)
s
Q
n
=g
Q
sin(
o
n +
Q
) +c
Q
. (6)
In (5) and (6), g
I
and g
Q
are gains,
I
and
Q
are initial phases,
and c
I
and c
Q
are dc offsets. These six parameters determine
the characteristics of the I/Q imbalance. They are all unknown
and have to be estimated from the receiver output {z
n
}.
B. I/Q Imbalance Parameters
We have the following denitions [13]. The gain imbalance
G is the quotient between the gain in the I branch and the gain
in the Q branch, i.e.,
G =
g
I
g
Q
. (7)
The quadrature skew Q is the phase difference relative to /2
between the two branches, i.e.,
Q =
I
Q
. (8)
The LO leakage L is the quotient between the total power of the
dc offset to the power of the sine waves, i.e.,
L = 2
c
2
I
+c
2
Q
g
2
I
+g
2
Q
(9)
where L is a scaled version of the leakage parameter that is
dened in [13]. The factor 2 in (9) is due to the power of the
sine waves, which are given by g
2
I
/2 and g
2
Q
/2, respectively.
Typically, G is measured in decibels, i.e., G
dB
= 20 log
10
G,
and L is measured in decibels referenced to 1 mW (dBm), i.e.,
L
dBm
= P +L
dB
, with P being the reference power in dBm.
Furthermore, the quadrature skew Q is measured in degrees,
i.e., Q
o
= 180 Q/. For a perfect I/Q demodulator, G
dB
=
0 dB, Q
o
= 0
, and L
dBm
= dB.
III. MEASURING THE RECEIVER I/Q IMBALANCE
A direct estimation of G, Q, and L from a set of recorded
baseband data {z
n
} is not feasible because of their nonlinear
relations to the observed data. Accordingly, we transform the
problem so that, rst, a more feasible set of parameters is
estimated. Once this initial set of estimates has been obtained,
estimates of the soughtfor G, Q, and Lare obtained by a proper
transformation.
A. I/Q Imbalance Parameters by Least Squares
The least squares estimate of the employed set of parameters
is given by the set that minimizes the least squares criterion
V (, ) =
N
n=1
z
n
s
n
(, )
2
(10)
where N denotes the number of complexvalued baseband
samples, and s
n
(, ) is a model of the receiver response
as a function of the parameter values. Here, the parameters
are gathered in a parameter vector and a scalar parameter
, where the latter parameter is the angular frequency of the
baseband representation of the excitation signal. One may argue
that parameter is known (at least up to some uncertainty)
and should not be estimated. For the considered USRPs, we
typically have some 10 to 20kHz frequency deviation in the
LO, implying that the accuracy of the estimate is expected
to be improved using as a free parameter. See the detailed
study of the problem in general in [18]. The soughtfor esti
mated parameter values and are given by the minimizing
argument
, = arg min
,
V (, ). (11)
The employed choice of a parameter vector is motivated
here. In view of (5) and (6), an approach would be to
use = (g
I
,
I
, c
I
, g
Q
,
Q
, c
Q
)
T
, where T denotes transpose.
One drawback with this parameterization is that the model
s
n
(, ) = s
I
n
(, ) +is
Q
n
(, ) becomes nonlinear in phases
I
and
Q
(and ), and, thus, nding the solution to (11)
becomes cumbersome. In the sequel, a complexvalued param
eterization, which is linear in its parameters, is employed.
B. Parameterization of the Signal Model
Consider s
n
(, ) = s
I
n
(, ) +is
Q
n
(, ), where the I and
Q components are given by (5) and (6), respectively. We seek
for an equivalent representation using a parameterization that is
linear in the parameters. Consider
s
n
(, ) = ae
in
+b
e
in
+c. (12)
In (12), a, b (or rather b
c
_
_
. (13)
Let a = a
r
+ia
i
and b = b
r
+ib
i
, where a
r
and a
i
are the
real and imaginary parts of a, etc. Relations (5) and (6) can
be rewritten as
s
I
n
(, )=
_
A
2
I
+B
2
I
sin
_
n+arctan 2(B
I
, A
I
)+
2
_
+c
I
(14)
s
Q
n
(, ) =
_
A
2
Q
+B
2
Q
sin (n + arctan 2(B
Q
, A
Q
)) +c
Q
(15)
where g
2
I
=A
2
I
+B
2
I
, g
2
Q
=A
2
Q
+B
2
Q
,
I
=arctan 2(B
I
, A
I
),
and
Q
= arctan 2(B
Q
, A
Q
). We relate the model given by
HNDEL AND ZETTERBERG: RECEIVER I/Q IMBALANCE: TONE TEST, SENSITIVITY ANALYSIS, AND THE USRP 707
(14) and (15) to (12) by
A
I
=a
r
+b
r
(16)
B
I
= (a
i
+b
i
) (17)
A
Q
=a
i
b
i
(18)
B
Q
=a
r
b
r
. (19)
Using the new sets of parameters, the gain imbalance G in (7)
and the quadrature skew Q in (8) can alternatively be written as
G =
_
A
2
I
+B
2
I
_
A
2
Q
+B
2
Q
=
a +b
a b
(20)
Q = arctan 2(B
I
, A
I
) arctan 2(B
Q
, A
Q
)
=
_
a +b
a b
_
(21)
where [] denotes the phase angle of the complexvalued
quantity within the brackets. Furthermore, with c = c
I
+ic
Q
,
the LO leakage in (9) can be written as
L = 2
c
2
I
+c
2
Q
A
2
I
+B
2
I
+A
2
I
+A
2
Q
=
c
2
a
2
+ b
2
. (22)
In summary, the I/Qimbalance problemhas been parameterized
in complexvalued parameters a, b, and c, which are gathered in
parameter vector , as in (13). The signal model that is tted to
the baseband data {z
1
, . . . , z
n
} is given by (12) and, using the
vector notation, can be written as s
n
(, ) = (e
in
, e
in
, 1).
The least squares estimation of and is treated here.
C. Least Squares Estimation
Consider N complexvalued baseband samples of the re
ceiver output and gather them in a column vector z, i.e.,
z =
_
_
z
1
.
.
.
z
N
_
_
. (23)
We seek the parameter values of and (which are denoted
by and , respectively) that are minimizing the least squares
criterion (10), where s
n
(, ) is the model output given by (12).
To rephrase the problem to vector notation, introduce the N 3
matrix C() as
C() =
_
_
e
i
e
i
1
.
.
.
.
.
.
.
.
.
e
iN
e
iN
1
_
_
. (24)
Then, the criterion in (10) can be rewritten as
V (, ) = (z C())
H
(z C()) (25)
where H denotes Hermitian transpose. The least squares prob
lem in (25) is a socalled separable least squares problem [19].
Accordingly, the problem can be solved in two steps. In the
rst step, parameter that enters the signal model C() in a
nonlinear fashion through C() is found by a nonlinear but 1D
search for the maximum of the condensed loss function [19]
V () = z
H
C()
_
C
H
()C()
_
1
C
H
()z. (26)
Once the maximizer to (26) is found, the estimate of the linear
parameters follows as the least squares solution to C( ) = z,
i.e.,
=
_
C
H
( )C( )
_
1
C
H
( )z. (27)
The aspects of numerical implementations of nonlinear search
methods to nd the maximum of (26) are beyond the scope
of this paper. In a related context, the maximization of (26)
around an initial value with the aid of the scalarbounded
nonlinear function minimization routine fminbnd provided by
MATLAB was considered in [20], which is an approach that
is employed in this paper as well. Another implementation is
provided by the socalled sevenparameter t in [21] (i.e., one
may note that and contain seven realvalued parameters in
total). One may note that similar estimation problems arise for
other engineering applications like impedance measurements
[22] and the estimation of particle size and velocity in laser
anemometry [23], [24]. The sevenparameter t and the method
in [23] are both nonlinear least squares types of methods
requiring some interpolation or iterations to provide the sought
after estimates. Noniterative alternative methods include the
correlationbased methods in [23] and the ellipsetting method
in [25] and [26], which is based on the work in [27]. From
Fig. 2, one may observe the spurious peaks that are not covered
by the present signal model, for example, the rst harmonics at
2F and its mirror distortion. An extension of the signal model
to cover harmonics is possible; see, for example, the different
approaches in [28][31] that cover generalizations of the IEEE
standardized fourparameter t [32][34]. A similar extension
of the present sevenparameter t is possible; however, it is
beyond the scope of this paper. The tradeoff between the gain
in performance and the complexity of the algorithm is expected
to be nonfavorable.
The aspects of nding a suitable initial value for numerical
optimization and estimation accuracy are provided here.
D. Finding an Initial Frequency Value
To nd a suitable initial value of the frequency estimate for
the maximization of (26), peak picking of the periodogram is a
natural approach. Consider the periodogram of data {z
n
}, i.e.,
P() =
1
N
n=1
z
n
e
in
2
=
1
N
z
T
c()
2
(28)
where c() is the rst column of C() in (24), i.e.,
c() =
_
_
e
i
.
.
.
e
iN
_
_
. (29)
708 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 59, NO. 3, MARCH 2010
TABLE I
NONLINEAR LEAST SQUARES FIT TO OBTAIN I/Q IMBALANCE
PARAMETERS FROM BASEBAND RECEIVER DATA
The periodogram shows peaks at =
o
, = 0, and =
o
. The heights of the peaks are approximately b
2
/N, c
2
/N,
and a
2
/N, respectively. Since the mirror distortion (contribu
tion at =
o
) and LO leakage ( = 0) are typically tens
of decibels below the peak at =
o
, it is natural to select the
location of the most dominant peak in the periodogram as an
initial estimate of the angular frequency.
To fully utilize the information contained in {z
n
}, we note
that the recorded response contains several periods of the
stimuli, i.e., N 2/
o
. For large values of N, the criterion
(26) can be simplied, noting that the product C
H
()C()
is approximately given by N I, where I is the unity ma
trix of size three. Accordingly, V () V
(), where V
()
denotes the largeN equivalent V
() = z
H
C()
2
/N, and
denotes an equality where only the dominant terms have
been retained. Decompose C() as C() = (c(), c
(), 1),
where the c() vectors are dened in (29). Then, V
() =
z
T
c()

2
+ z
T
c()
2
+ z
T
1
2
, where 1 is the vector of
ones. The third term is independent of and can be omitted in
the maximization. To summarize, in a large N, the least squares
solution is given by
0
= arg max
z
T
c()
2
+
z
T
c()
2
N
. (30)
We observe that the frequency estimate is given by the location
of the maxima of the sum of periodogram P() and its mirror
P(). Because of the symmetry, it is sufcient to search
for the peak in the interval (0, ). This approach to nd an
initial value
0
is employed in the numerical examples, with
the periodogram calculated using the fast Fourier transform of
four to eighttimezeropadded data.
The least squares modeling of I/Q imbalance parameters is
summarized in Table I.
E. Accuracy Aspects
To make the analysis feasible, we rely on a Gaussian assump
tion, which means that the noise terms v
I
n
and v
Q
n
in (3) and
(4) are jointly Gaussian. That is, v
n
= v
I
n
+iv
Q
n
is a circular
complex Gaussian with variance 2
2
. We have the following
results on estimation accuracy.
Theorem 1: Consider any unbiased estimator producing the
estimates
G,
Q, and
L based on the receiver output z given
by (23). The entries {z
n
} of z fulll z
n
= s
n
(, ) +v
n
, with
s
n
(, ) being dened by (12) and v
n
being a circular complex
Gaussian with variance 2
2
. If the number of samples N
2/
o
, then the asymptotic variances of the estimates
G,
Q,
and
L are bounded from below by
2
G
=
1 +G
2
N SNR
Q
(31)
2
Q
=
1 +G
2
N SNR
I
(32)
2
L
=
2L(1 +L)
N SNR
(33)
where the SNRs in the I (SNR
I
) and Q (SNR
Q
) channels are
given by
SNR
I
=
g
2
I
2
2
SNR
Q
=
g
2
Q
2
2
. (34)
Furthermore, SNR denotes the average SNR, i.e., SNR =
(SNR
I
+ SNR
Q
)/2.
Proof: The lower bound is given by the CramrRao
bound (CRB) [19]. The details for the considered signal model
are provided in the Appendix IA.
Since gain Gis typically close to unity and L is close to zero,
the rule of thumb directly follows as
2
G
=
2
Q
= 2/(N SNR)
and
2
L
= 2L/(N SNR) since G = 1 implies equal channel
SNRs. If the resolution of the ADCs is the bottleneck regarding
noise powerthat is, quantization noise is the major noise
source and can be assumed Gaussianthe noise variance for
an ADC full scale within [1, 1] is given by
2
=
2
2B
3
(35)
where B is the ADC resolution in number of bits.
IV. EXPERIMENTAL EVALUATION
A. Simulation Example
We consider a scenario given by G
dB
= 1.0 dB, Q
o
=
1.0
, and L
dB
= 40.0 dB. The additive noise is white and
Gaussian, with an average SNR of 74 dB, corresponding to
the signaltonoiseanddistortion ratio of B = 12 bits ADCs.
Only N = 32 samples are used, with the angular frequency set
to
o
= 0.15, that is, coherent sampling is not employed. Thus,
the receiver output z
n
contains approximately ve periods. In
Fig. 3, the results are shown based on 10 000 independent
simulation runs. The gure displays the empirical histograms
of the estimates that are compared with the probability density
function of a Gaussian distribution with variance given by
Theorem 1 and centered around the parameter of interest.
From the results in Fig. 3, we note an excellent agreement
between simulation results and predictions made from Theorem
1. In fact, under the Gaussian assumption, the method of max
imum likelihood is given by the minimizer of the least squares
criterion (10) [20], [24]. The performance of and in terms
of error variance is expected to (at least as N ) coincide
HNDEL AND ZETTERBERG: RECEIVER I/Q IMBALANCE: TONE TEST, SENSITIVITY ANALYSIS, AND THE USRP 709
Fig. 3. Simulation results based on 10 000 independent runs. (Circles) Normalized histograms with ten bins of the estimated
G
dB
,
Q
o
, and
L
dB
. In comparison,
the (solid lines) Gaussian probability density functions N(G,
2
G
), N(Q,
2
Q
), and N(L,
2
L
) are shown. The vertical dashed lines indicate the center of the
Gaussian distributions, that is, the location of the true parameters G
dB
= 1.0 dB, Q
o
= 1.0
, and L
dB
= 40 dB.
TABLE II
DATASHEET VALUES FOR AD8347 AT 1.905 GHz
with the CRB. Accordingly, by the invariance principle of the
method of maximum likelihood, the asymptotic variance of the
soughtfor imbalance parameters is given by Theorem 1.
B. I/Q Imbalance of Sample USRPs
The FLEXRF family of daughterboards, which are widely
used with the USRP platform, is designed around the AD8347
directconversion quadrature demodulator. Its specication re
garding the imbalance is given in Table II.
With reference to Fig. 4, the USRP receivers were excited by
an F
RF
= 882.3 MHz sine wave at 12 dBm by an HP8656B,
followed by a MiniCircuits FK3000 frequency doubler, a Mini
Circuits SHP900 highpass lter, MiniCircuits attenuators
(3 10 dB and 2 5 dB), and a ZAPD30 splitter for the
simultaneous excitation of the two RF inputs of the USRP. The
measured input signal level was P
RF
= 60 dBm. No internal
downconversion was employed in the USRP. For each USRP
receiver, a sequence of 100 000 of baseband data was collected.
The sampling rate of the USRP ADCs was F
S
= 4 MHz,
followed by digital ltering and subsampling. With F
LO
=
1764 MHz, we end up with normalized angular frequency
o
= 0.15.
In the rst experiment, the baseband output from a USRP
receiver was divided into nonoverlapping blocks of length
N = 64 samples. The imbalance test according to Table I was
employed. The results based on the statistics employing 1500
nonoverlapping segments of data are shown in Fig. 5. As a
reference point, the average parameter (based on the results
from the independent runs) values were calculated to illustrate
the results in Theorem 1. We may note a good agreement
between the experimental results and the theoretical predictions
for the gain imbalance and the quadrature skew. We also
note that three out of the four receivers have performance
according to the datasheet of the AD8347 directconversion
quadrature demodulator, whereas the fourth unit has outlier
performance. The outlier performance was isolated to one of
the FLEXRF1800 daughterboards.
For the LO leakage, the theoretical predictions do not fully
describe the experimental results, as shown in Fig. 6. By
Theorem 1, the L estimates are described as Gaussian with
mean value L and variance
2
L
. However, by construction, the
estimates are positive and given by the transformation (22).
This inherent positiveness, in combination with small values
of
L when compared with its standard deviation, results in that
part of the Gaussian distribution corresponding to negative L
values, which is folded around L = 0. The inexactness of the
theoretical predictions may, thus, be explained by the folded tail
of the Gaussian distribution that is added to the tail for positive
L values. The LO leakage of the USRP is further examined.
C. LO Leakage of the USRP
As mentioned earlier, the positiveness of
L results in the
folding of the Gaussian bell around zero. This folding adds a
bias to the estimate, that is,
L in the mean provides too high
a value for the LO leakage. An increase in the number of data
yields that the probability density function around its mean is
narrowed and approaches a Gaussian shape. Accordingly, the
710 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 59, NO. 3, MARCH 2010
Fig. 4. Block diagram of measurement testbed for simultaneous excitation of the two channels of the USRP device under test. Measured input power
P
RF
= 60 dBm.
Fig. 5. USRP results based on 1500 nonoverlapping segments, each of length N = 64. Histograms of the estimated
G
dB
and
Q
o
. In comparison, the
(solid lines) Gaussian probability density functions N(
G,
2
G
) and N(
Q,
2
Q
) are shown, where the bar indicates that a parameter value is calculated as an
average over all available data. USRP #1: () Channel 1; () channel 2. USRP #2: () Channel 1; () channel 2.
bias of the estimate of LO leakage is decreased. When N is
sufciently large (i.e., N is such that L
L
), the bias is
negligible, yielding unbiased estimates of the LO leakage. To
investigate this property, the average estimate is calculated, i.e.,
L =
1
M
M
m=1
L
m
(36)
where M is the number of nonoverlapping segments of the
data, and
L
m
is the LO leakage estimate based on the data in
segment m. A sequence of 100 000 of baseband data from one
of the USRP receivers is employed, where M is set to M = 100
for sequences of length up to N = 512 samples. For higher
values of N, the number of segments is given by the maximum
number of nonoverlapping segments provided by the 100 000
data sequence.
The results that are displayed in Fig. 7 show the average
estimated L value, as well as the outcome from each of the
data sets to illustrate the spread. One may note that the estimate
attens out (in the mean) around 74 dB, where the knee is
given by N = 8192 samples. One may also note that using
N = 64 samples indicates a bias (in the mean) of 16 dB and
a spread in the range from 50 to 80 dB.
To compare the obtained LO leakage values with the
datasheet gures provided in Table II, we note that the AD8347
input power at RFIP is equal to P
RFIP
= P
RF
+ 13 dBm,
taking the gain of the FLEXRF1800internal MGA82563 am
plier into account. Accordingly, the leakage reads L
dBm
=
47 +L
dB
dBm at the RFIP of the AD8377. Measured L
dB
values for the considered USRPs are in the range from 60 to
70 dBm, yielding leakage values that are below 100 dBm
at RFIP, because of the enabled internal automatic dc offset
calibration provided by the USRP. Additional measurements
indicate that disabling the dc compensation yields leakage
values in the range from 66 to 90 dBm at RFIP, which
is in accordance with the datasheet values. Disabling this
feature does not inuence the measured gain imbalance or
quadrature skew.
HNDEL AND ZETTERBERG: RECEIVER I/Q IMBALANCE: TONE TEST, SENSITIVITY ANALYSIS, AND THE USRP 711
Fig. 6. USRP results based on 1500 nonoverlapping segments, each of length N = 64. Histograms of the estimated
L
dB
. In comparison, the (solid line)
Gaussian probability density function N(
L,
2
L
) is shown, where the bar indicates that a parameter value is calculated as an average over all available data.
USRP #1: () Channel 1; () channel 2. USRP #2: () Channel 1; () channel 2.
Fig. 7. USRP results based on nonoverlapping segments of length N. The
gure displays the estimated LO leakage as a function of N, where each
outcome is indicated by a dot. (Solid line) Average estimated LO leakage. The
full batch of data is 100 000 samples, and the number of outcomes per given
value of N varies with the segment length.
V. CONCLUSION
We have proposed a least squares approach to determine the
I/Q imbalance of a directconversion receiver. By reparameter
ization of the problem, a least squares problem with six real
valued parameters that enter the problem in a linear fashion
is obtained, with only the angular frequency of the baseband
stimuli entering the problem in a nonlinear fashion. The method
has been summarized in Table I.
Under a Gaussian assumption, the accuracy of the method
has been addressed, yielding the closedform results in
Theorem 1. The test method for I/Q imbalance has been applied
to simulated data. In the simulation study, we have employed
a signal that covered ve periods, still showing an excellent
agreement between theoretical predictions and simulation re
sults. Experimental data from four different USRP receivers
have been collected. It has been shown that gain imbalance and
quadrature skeware accurately estimated by employing data that
cover only a handful of full periods of the test stimuli, which
highlights the practical relevance of the derived test method.
It has also been shown that estimating the LO leakage is a
more complicated problem, not because of the accuracy aspects
but because of the systematic errors, providing errors on the
order of 1020 dB in scenarios where gain imbalance and
quadrature skew are accurately estimated. From Theorem 1,
the origin of this systematic error is readily explained by the
folding of the distribution of the estimates around zero. The
effect of folding of the Gaussian bell is well known [35], and,
accordingly, the systematic error may be predicted.
By examining four FLEXRF1800equipped USRPs, we have
not only validated the AD8347 datasheet performance regard
ing gain imbalance, quadrature skew, and LO leakage in three
out of four cases but also observed the outlier performance of
a sample FLEXRF1800 daughterboard. We have found that the
automatic dc offset calibration of the USRP is effective. The
evaluation of USRP receivers is an interest of its own, owing to
the widespread use of the USRPs for research, education, and
development.
APPENDIX I
A. Proof of Theorem 1
There are a total of seven realvalued parameters to be
estimated. Decomposing the model s
n
(, ) in (12) into its real
and imaginary counterparts yields (14) and (15), which, in turn,
may be rewritten as
s
I
n
() =A
I
cos
_
n +
2
_
+B
I
sin
_
n +
2
_
+c
I
(37)
s
Q
n
() =A
Q
cos(n) +B
Q
sin(n) +c
Q
. (38)
712 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 59, NO. 3, MARCH 2010
Here, is introduced as the sevenparameter vector
= (A
I
, B
I
, c
I
, , c
Q
, B
Q
, A
Q
)
T
. (39)
The CRB is an objective bound on the achievable estimation
error variance of any method [19]. The bound corresponding
to the considered model parameterized in in (39) has been
recently derived in [20]. For further information on the topic,
see [18] and [36][38]. The bound is given by the inverse of the
Fisher information matrix J() [19]
J() = E
_
_
ln p(z; )
__
ln p(z; )
_
T
_
(40)
where p(z; ) denotes the probability density function, and the
derivative is evaluated at the true parameters. Derivations of
J() are provided in [20]. To derive tractable expressions, an
assumption of large N is employed. In practice, the assump
tion corresponds to a signal generator stimulus that covers
much more than a full period of the sine wave. Under this
assumption, an analytic expression for the inverse J()
1
of
the Fisher information matrix is provided in [20]. The result in
[20, eq. (17)] is not directly applicable to the problem statement
that is considered herein. With h = A
2
I
+B
2
I
+A
2
Q
+B
2
Q
=
2(a
2
+ b
2
), the result in terms of all entries in the resulting
J
1
()
1
is reported in (61), shown at the bottom of the page.
The diagonal elements of (61) correspond to the lower bound
on the error variance of the corresponding parameters in . To
transform the results, we introduce parameter vector
= (G, Q, L, , c
Q
, B
Q
, A
Q
). (41)
We have the following general result [19]:
J()
1
=
_
_
J()
1
_
_
T
(42)
where the (k, )th entry of [/] is
k
/
. We have to
derive the following derivatives:
G
A
I
=
A
I
_
(A
2
I
+B
2
I
)
_
A
2
Q
+B
2
Q
_
(43)
G
B
I
=
B
I
_
(A
2
I
+B
2
I
)
_
A
2
Q
+B
2
Q
_
(44)
G
B
Q
=
B
Q
A
2
Q
+B
2
Q
_
A
2
I
+B
2
I
_
A
2
Q
+B
2
Q
(45)
G
A
Q
=
A
Q
A
2
Q
+B
2
Q
_
A
2
I
+B
2
I
_
A
2
Q
+B
2
Q
(46)
Q
A
I
=
B
I
A
2
I
+B
2
I
(47)
Q
B
I
=
A
I
A
2
I
+B
2
I
(48)
Q
B
Q
=
A
Q
A
2
Q
+B
2
Q
(49)
Q
A
Q
=
A
Q
A
2
Q
+B
2
Q
(50)
L
A
I
=
4A
I
_
c
2
I
+c
2
Q
_
_
A
2
I
+B
2
I
+A
2
Q
+B
2
Q
_
2
(51)
L
B
I
=
4B
I
_
c
2
I
+c
2
Q
_
_
A
2
I
+B
2
I
+A
2
Q
+B
2
Q
_
2
(52)
L
C
r
=
4c
I
A
2
I
+B
2
I
+A
2
Q
+B
2
Q
(53)
L
C
i
=
4c
Q
A
2
I
+B
2
I
+A
2
Q
+B
2
Q
(54)
L
B
Q
=
4B
Q
_
c
2
I
+c
2
Q
_
_
A
2
I
+B
2
I
+A
2
Q
+B
2
Q
_
2
(55)
L
A
Q
=
4A
Q
_
c
2
I
+c
2
Q
_
_
A
2
I
+B
2
I
+A
2
Q
+B
2
Q
_
2
. (56)
All other derivatives are equal to zero. Inserting (43)(56) in
(42) using the J()
1
given by (61), a straightforward but
tedious calculation results in
J()
1
=
_
_
_
_
_
_
_
_
_
_
2
G
0 0 0 0 x x
0
2
Q
0 0 0 x x
0 0
2
L
0 x x x
0 0 0
2
0 x x
0 0 x 0
2
C
i
0 0
x x x x 0
2
B
Q
x
x x x x 0 x
2
A
Q
_
_
_
_
_
_
_
_
_
_
(57)
where x and x denote generic nonzero elements, where
the different values of the covariances x can be found in
J
1
()
1
=
2
2
Nh
_
_
_
_
_
_
_
_
_
_
_
h + 3B
2
I
3B
I
A
I
0
6B
I
N
0 3B
I
A
Q
3B
I
B
Q
3B
I
A
I
h + 3A
2
I
0
6A
I
N
0 3A
I
A
Q
3A
I
B
Q
0 0
h
2
0 0 0 0
6B
I
N
6A
I
N
0
12
N
2
0
6A
Q
N
6B
Q
N
0 0 0 0
h
2
0 0
3B
I
A
Q
3A
I
A
Q
0
6A
Q
N
0 h + 3A
2
Q
3B
Q
A
Q
3B
I
B
Q
3A
I
B
Q
0
6B
Q
N
0 3
Q
A
Q
h + 3B
2
Q
_
_
_
_
_
_
_
_
_
_
_
(61)
HNDEL AND ZETTERBERG: RECEIVER I/Q IMBALANCE: TONE TEST, SENSITIVITY ANALYSIS, AND THE USRP 713
[20, eq. (17)]. Explicit values of the covariances marked by an
x are reported elsewhere. The resulting
2
G
,
2
Q
, and
2
L
are
given by
2
G
=
2
2
N
A
2
I
+B
2
I
+A
2
Q
+B
2
Q
_
A
2
Q
+B
2
Q
_
2
(58)
2
Q
=
2
2
N
A
2
I
+B
2
I
+A
2
Q
+B
2
Q
_
A
2
Q
+B
2
Q
_
(A
2
I
+B
2
I
)
(59)
2
L
=
16
2
N
_
C
2
r
+C
2
i
_ _
2
_
C
2
r
+C
2
i
+A
2
I
+B
2
I
+A
2
Q
+B
2
Q
_
_
A
2
I
+B
2
I
+A
2
Q
+B
2
Q
_
3
.
(60)
Quantities
2
,
2
C
i
,
2
B
Q
, and
2
A
Q
can be found in
[20, eq. (17)]; note the slightly different denition of h in [20]
and that the derivation in [20] is slightly more general since the
noise variances in the two channels are not necessarily equal.
Employing the denition of G in (7),
2
G
and
2
Q
in (31) and
(32), respectively, directly follow. In a similar vein, (33) follows
by employing the denition of L in (9).
ACKNOWLEDGMENT
The authors would like to thank, in general, the GNURADIO
initiative (http://www.gnuradio.org) for their helpful guidance
on the USRP and, in particular, M. Ettus for providing valuable
input on the content of an earlier version of this paper.
REFERENCES
[1] G. Fettweis, M. Lhning, D. Petrovic, M. Windisch, P. Zillmann, and W.
Rave, Dirty RF: A new paradigm, Int. J. Wirel. Inf. Netw., vol. 14, no. 2,
pp. 133148, Jun. 2007.
[2] M. Valkama, M. Renfors, and V. Koivunen, Advanced methods for
I/Q imbalance compensation in communication receivers, IEEE Trans.
Signal Process., vol. 49, no. 10, pp. 23352344, Oct. 2001.
[3] J. Tubbax, B. Come, L. Van der Perre, S. Donnay, M. Engels,
H. De Man, and M. Moonen, Compensation of IQ imbalance and phase
noise in OFDM systems, IEEE Trans. Wireless Commun., vol. 4, no. 3,
pp. 872877, May 2005.
[4] L. Anttila, M. Valkama, and M. Renfors, Circularitybased I/Qimbalance
compensation in wideband directconversion receivers, IEEE Trans. Veh.
Technol., vol. 57, no. 4, pp. 20992113, Jul. 2008.
[5] P. F. Morlat, J. C. N. Perez, G. Villemaud, J. Verdier, and J. M. Gorce,
On the compensation of RF impairments with multiple antennas in SIMO
OFDM systems, in Proc. IEEE 64th Veh. Technol. Conf., Montreal, QC,
Canada, Sep. 2528, 2006, pp. 15.
[6] H. Lundin, P. Svedman, X. Zhang, M. Skoglund, P. Hndel, and
P. Zetterberg, ADC imperfections in multiple antenna wireless
systemsAn experimental study, in Proc. 9th Eur. Workshop ADC
Modelling Testing, Athens, Greece, Sep. 2004, pp. 808813.
[7] X. Li and M. Ismail, MultiStandard CMOS Wireless Receivers: Analysis
and Design. Norwell, MA: Kluwer, 2002.
[8] M. Valkama, J. Pirskanen, and M. Renfors, Signal processing challenges
for applying software radio principles in future wireless terminals: An
overview, Int. J. Commun. Syst., vol. 15, no. 8, pp. 741769, Oct. 2002.
[9] S. Cass, Hardware for your software radio, IEEE Spectr., vol. 43, no. 10,
pp. 5156, Oct. 2006.
[10] J. McEllroy, J. F. Raquet, and M. A. Temple, Use of a software radio
to evaluate signals of opportunity for navigation, in Proc. 19th Int. Tech.
Meeting Satellite Division, ION GNSS, 2006, vol. 1, pp. 126133.
[11] Z. Yan, Z. Ma, H. Cao, G. Li, and W. Wang, Spectrum sensing, access
and coexistence testbed for cognitive radio using USRP, in Proc. 4th
IEEE Int. Conf. Circuits Syst. Commun., May 2628, 2008, pp. 270274.
[12] R. Dhar, G. George, A. Malani, and P. Steenkiste, Supporting integrated
MAC and PHY software development for the USRP SDR, in Proc. 1st
IEEE Workshop Netw. Technol. Softw. Dened Radio Netw., Sep. 2525,
2006, pp. 6877.
[13] G. Zoka, Rened I/Q: Imbalance measurements, Microw. RF, vol. 43,
no. 6, pp. 7283, 2004.
[14] L. Angrisani, I. Ghidini, and M. Vadursi, A new method for I/Q im
pairments detection and evaluation of OFDM transmitters, IEEE Trans.
Instrum. Meas., vol. 55, no. 5, pp. 14801486, Oct. 2006.
[15] L. Angrisani, A. Napolitano, and M. Vadursi, Measuring I/Q impair
ments in WiMAX transmitters, IEEE Trans. Instrum. Meas., vol. 58,
no. 5, pp. 12991306, May 2009. DOI: 10.1109/TIM.2008.2009141.
[16] K. Asami, An algorithm to evaluate wideband quadrature mixers, in
Proc. IEEE Int. Test Conf., Oct. 2126, 2007, pp. 17.
[17] E. S. Erdogan and S. Ozev, Singlemeasurement diagnostic test method
for parametric faults of I/Q modulating RF transceivers, in Proc. 26th
IEEE VLSI Test Symp., Apr. 27May 1, 2008, pp. 209214.
[18] T. Andersson and P. Hndel, IEEE Standard 1057, CramrRao bound
and the parsimony principle, IEEE Trans. Instrum. Meas., vol. 55, no. 1,
pp. 4453, Feb. 2006.
[19] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
Theory, vol. 1. Upper Saddle River, NJ: PrenticeHall, 1993.
[20] P. Hndel, Parameter estimation employing a dual channel sinewave
model under a Gaussian assumption, IEEE Trans. Instrum. Meas.,
vol. 57, no. 8, pp. 16611669, Aug. 2008.
[21] P. M. Ramos and A. C. Serra, A new sinetting algorithm for accu
rate amplitude and phase measurements in two channel acquisitions,
Measurement, vol. 41, no. 2, pp. 135143, Feb. 2008.
[22] P. M. Ramos, M. Fonseca da Silva, and A. C. Serra, Low frequency
impedance measurement using sinetting, Measurement, vol. 35, no. 1,
pp. 8996, Jan. 2004.
[23] P. Hndel and A. HstMadsen, Estimation of velocity and size of par
ticles from two channel laser anemometry measurements, Measurement,
vol. 21, no. 3, pp. 113123, Jul. 1997.
[24] A. HstMadsen and K. Andersen, Lower bounds for estimation of fre
quency and phase of Doppler signals, Meas. Sci. Technol., vol. 6, no. 6,
pp. 637652, Jun. 1995.
[25] F. M. Janeiro, P. M. Ramos, M. Tlemcani, and A. C. Serra, Analysis of
an noniterative algorithm for the amplitude and phase difference estima
tion of two acquired sinewaves, in Proc. XVIII IMEKO World Congr.,
Rio de Janeiro, Brazil, Sep. 2006.
[26] P. M. Ramos, F. M. Janeiro, M. Tlemcani, and A. C. Serra, Uncertainty
analysis of impedance measurements using DSP implemented ellipse t
ting algorithms, in Proc. IEEE IMTC, Victoria, BC, Canada, May 1215,
2008, pp. 463467.
[27] A. Fitzgibbon, M. Pilu, and R. B. Fisher, Direct least square tting of
ellipses, IEEE Trans. Pattern Anal. Mach. Intell., vol. 21, no. 5, pp. 476
480, May 1999.
[28] T. Andersson and P. Hndel, Multipletone estimation by IEEE Standard
1057 and the expectationmaximization algorithm, IEEE Trans. Instrum.
Meas., vol. 54, no. 5, pp. 18331839, Oct. 2005.
[29] T. Andersson and P. Hndel, Toward a standardized multisinewave t
algorithm, in Proc. 9th Eur. Workshop ADC Modelling Testing, Athens,
Greece, Sep. 2001, pp. 337342.
[30] P. M. Ramos, M. Fonesca da Silva, R. C. Martins, and A. C. Serra, Sim
ulation and experimental results of multiharmonic leastsquares tting
algorithms applied to harmonic signals, IEEE Trans. Instrum. Meas.,
vol. 55, no. 2, pp. 646651, Apr. 2006.
[31] P. M. Ramos and A. C. Serra, Least squares multiharmonic tting:
Convergence improvements, IEEE Trans. Instrum. Meas., vol. 56, no. 4,
pp. 14121418, Aug. 2007.
[32] IEEE Standard for Digitizing Waveform Recorders, 2008. IEEE Standard
1057, IEEE Standard, 10571994.
[33] IEEE Standard for Terminology and Test Methods for AnalogtoDigital
Converters, 2001. IEEE Standard, 12412000.
[34] T. E. Linnenbrink, W. B. Boyer, N. G. Paulter, Jr., and S. J. Tilden, IEEE
TC10: Update 2006, in Proc. IEEE Instrum. Meas. Technol. Conf.,
Sorrento, Italy, Apr. 2006, pp. 130133.
[35] F. C. Leone, R. B. Nottingham, and L. S. Nelson, The folded normal
distribution, Technometrics, vol. 3, no. 4, pp. 543550, 1961.
[36] P. Hndel, Properties of the IEEESTD1057 fourparameter sine wave
t algorithm, IEEE Trans. Instrum. Meas., vol. 49, no. 6, pp. 11891193,
Dec. 2000.
[37] D. Rife and R. Boorstyn, Single tone parameter estimation from discrete
time observations, IEEE Trans. Inf. Theory, vol. IT20, no. 5, pp. 591
598, Sep. 1974.
[38] A. Nehorai and B. Porat, Adaptive comb ltering for harmonic
signal enhancement, IEEE Trans. Acoust., Speech, Signal Process.,
vol. ASSP34, no. 5, pp. 11241138, Oct. 1986.
714 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 59, NO. 3, MARCH 2010
Peter Hndel (S88M94SM98) received the
Ph.D. degree from Uppsala University, Uppsala,
Sweden, in 1993.
From 1987 to 1993, he was with Uppsala Uni
versity. From 1993 to 1997, he was with Ericsson
AB, Kista, Sweden. From 1996 to 1997, he was also
with Tampere University of Technology, Tampere,
Finland. Since 1997, he has been with the Royal In
stitute of Technology, Stockholm, Sweden, where he
is currently a Professor of signal processing with the
Signal Processing Laboratory, ACCESS Linnaeus
Center. From2000 to 2006, he was with the Swedish Defence Research Agency.
He is currently a Guest Professor with the University of Gvle, Gvle, Sweden.
He has served as a member of the editorial board of the EURASIP Journal on
Advances in Signal Processing. He has also served as a member of the editorial
advisory board of Recent Patents on Electrical Engineering. He is a member
of the editorial board of Hindawis Research Letters in Signal Processing and
Journal of Electrical and Computer Engineering.
Dr. Hndel has served as an Associate Editor for the IEEE TRANSACTIONS
ON SIGNAL PROCESSING.
Per Zetterberg was born in Uppsala, Sweden, in
1968. He received the M.S. degree in electrical en
gineering from Lule University, Lule, Sweden, in
1993 and the Ph.D. degree in electrical engineer
ing from the Royal Institute of Technology (KTH),
Stockholm, Sweden, in 1997.
He then joined Radio Design AB, where he
worked with smart antenna research and implemen
tation. Since 2002, he has been with Signal Process
ing Laboratory, ACCESS Linnaeus Center, Royal
Institute of Technology, working full time with radio
propagation and testbeds. He is or has been involved in the EU projects
TSUNAMI II, NEWCOM, ACE, WINNER, WINNER II, and COOPCOM.